A method for early warning of sneezing and respiratory disease in broilers based on CNN-SE parameter fusion

Abstract

This invention discloses a broiler sneezing respiratory disease early warning method based on CNN-SE parameter fusion, belonging to the field of livestock and poultry breeding environment monitoring technology. The method includes establishing an infection model through challenge with *Avianella paragallinarum*, collecting and preprocessing full-time-series audio data; constructing a multi-dimensional structured feature map, training the CNN+SE model, and identifying three types of sounds: sneezing, calling, and other sounds; calculating a comprehensive score for each hourly audio segment by selecting specific intensity and spectral acoustic features and combining them with the chicken call frequency and sneezing frequency; determining a comprehensive threshold using two thresholding methods, and setting a three-level early warning discrimination method. Therefore, the above-mentioned CNN-SE parameter fusion-based broiler sneezing respiratory disease early warning method effectively improves the ability to distinguish respiratory sneezing sounds, significantly enhances the sensitivity and stability of the system, avoids over-alarms and under-alarms, and is suitable for early warning of respiratory diseases with sneezing as the main symptom.

Description

Technical Field

[0001] This invention relates to the field of livestock and poultry breeding environment monitoring technology, and in particular to a method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion. Background Technology

[0002] In modern intensive chicken farms, respiratory diseases such as infectious rhinitis and mycoplasma infection are common, spread rapidly, and cause great harm. Their early clinical manifestations are often acoustic abnormalities such as sneezing and coughing. Collecting the behavioral sounds of chickens has become a low-invasive, high-frequency source of information.

[0003] However, current chicken health monitoring mainly relies on manual inspections or traditional environmental monitoring equipment, lacking an automatic identification and early quantitative warning system for chicken behavioral acoustic signals. Furthermore, although some studies have adopted sound recognition-based disease detection methods, their training data largely comes from actual chicken farms, lacking experimental label validation, and the models are highly dependent on external models and have poor versatility. Summary of the Invention

[0004] The purpose of this invention is to provide an early warning method for sneezing and respiratory diseases in broilers based on CNN-SE parameter fusion, so as to solve the problems of difficulty in identifying abnormal sounds in large-scale caged chickens and the lack of clear threshold warning standards.

[0005] To achieve the above objectives, this invention provides a method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion, comprising the following steps:

[0006] S1. An infection model was established by challenging with Avianella paragallinarum, and full-time-series audio data was collected. The collected audio data was then preprocessed.

[0007] S2. Determine the acoustic feature parameters and concatenate them in matrix form to construct a multi-dimensional structured feature map as the model input. At the same time, use the CNN+SE model to construct a three-classification model to identify sneezing, crowing and other sound types in chicken sounds.

[0008] S3. Based on the trained CNN+SE model, classify and identify the audio data for each hour and obtain acoustic features of intensity and spectrum. Combine this with the frequency of chicken calls and sneezing to calculate the comprehensive score for each hour.

[0009] S4. By using the maximum value and standard deviation margin method during the healthy period and the optimal point method based on the ROC curve of infection label, a comprehensive threshold is determined and a three-level early warning level discrimination method is set to provide early warning for chicken diseases.

[0010] Furthermore, in S1, the preprocessing operation includes resampling and denoising; the denoising process includes suppressing background noise by using spectral gating, performing frame processing on the audio, enhancing the high-frequency response by pre-emphasis, smoothing the edges of each frame using a Hamming window, and extracting the acoustic signal of the active segment using short-time energy and short-time zero-crossing rate.

[0011] Furthermore, in S2, the acoustic characteristic parameters include the overall envelope, Mel frequency cepstral coefficients, Mel subband energy distribution, chromaticity spectrum, and spectral contrast.

[0012] Furthermore, in S2, the CNN+SE model includes a local feature extraction network and a channel attention module; the local feature extraction network is based on CNN and consists of convolutional layers, batch normalization, and activation functions; the channel attention module is used to enhance the ability to perceive channel weights.

[0013] Furthermore, in S3, the acoustic features of intensity include the average root mean square energy, the root mean square energy standard deviation, the maximum loudness, and the maximum amplitude; the acoustic features of spectrum include the spectral centroid and the spectral flow.

[0014] Furthermore, S3 includes normalization processing of acoustic features of intensity and spectrum, as well as chicken call and sneeze frequencies.

[0015] Furthermore, in S3, the expression for the composite score is as follows:

[0016] ;

[0017] In the formula, Indicates the first The overall score for the hour; Normalized numerical values ​​representing the frequency of sneezes; Represents the normalized values ​​of acoustic parameters; This represents the weight of the corresponding parameter, and satisfies... .

[0018] Furthermore, S4 includes determining the first threshold using the maximum value and standard deviation margin method during the healthy period, determining the second threshold using the optimal point method based on the ROC curve of the infection label, and determining the comprehensive threshold using a weighted average, as shown in the following expression:

[0019] ;

[0020] in,

[0021] ;

[0022] ;

[0023] In the formula, As a comprehensive threshold, The first threshold, The second threshold, To determine the degree of confidence in the ROC curve, This represents the maximum overall score for each hour of audio recordings during the health period. This represents the standard deviation of the overall score for each hour of audio recordings taken during the healthy period. This is the margin coefficient. Indicates the true rate, This indicates the false positive rate.

[0024] Furthermore, S4 includes updating thresholds using monitoring data during the actual production process.

[0025] Therefore, the broiler sneezing respiratory disease early warning method based on CNN-SE parameter fusion described above has the following technical effects:

[0026] (1) The present invention adopts a deep neural network structure of CNN+SE model, combined with multi-feature fusion input, which can extract key features of abnormal sounds such as sneezing and calling of chickens more comprehensively than traditional models, effectively improving the model's recognition accuracy of abnormal respiratory sounds.

[0027] (2) This invention combines acoustic features with sneezing frequency and chicken call frequency, which can better reflect the acoustic changes of abnormal breathing and has a stronger discrimination ability, adapting to the actual needs of early warning of respiratory diseases in broilers; at the same time, by comprehensively determining the three-level early warning threshold through ROC curve and the maximum value + standard deviation margin of the healthy period, the sensitivity and stability of the system are significantly enhanced, avoiding excessive alarms and missed alarms.

[0028] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0029] Figure 1 This is a flowchart of a broiler sneezing and respiratory disease early warning method based on CNN-SE parameter fusion;

[0030] Figure 2 This is the training process of the CNN-SE model in an embodiment of a broiler sneezing and respiratory disease early warning method based on CNN-SE parameter fusion. Detailed Implementation

[0031] The present invention will be explained in more detail through the following embodiments. The purpose of disclosing the present invention is to protect all changes and modifications within the scope of the present invention. The present invention is not limited to the following embodiments.

[0032] like Figure 1As shown, this invention provides a method for early warning of sneezing and respiratory diseases in broilers based on CNN-SE parameter fusion. An infection model is established through challenge with Avianella paragallinarum, high-quality audio is collected throughout the process, and a three-classification model is constructed using the CNN+SE (Squeeze-and-Excitation) module to identify sneezing, croaking, and other sound types in chicken sounds. At the same time, with numerous acoustic parameters, representative parameters of intensity and spectrum are systematically screened, and combined with two behavioral statistical indicators, sneezing frequency and croaking frequency, a comprehensive scoring system based on weighted fusion is established for early warning of chicken diseases. Sneezing frequency serves as the dominant indicator to reflect the degree of respiratory abnormality, croaking frequency provides auxiliary behavioral reference, and six types of acoustic features capture underlying physical changes. The three work together to construct a comprehensive representation of the disease state. In addition, this embodiment introduces two early warning threshold calculation strategies: (1) the maximum value during the healthy period + standard deviation margin method; (2) the optimal point method based on the ROC curve of infection label. By combining the results of the two methods, with sneezing frequency set as the dominant parameter (weight 50%) and the remaining seven parameters used as auxiliary weights to generate a comprehensive score S(t), and setting three levels of boundaries: green (normal), yellow (warning), and red (abnormal), this method not only improves the sensitivity of the warning but also enhances its interpretability.

[0033] (1) Audio data acquisition and preprocessing:

[0034] First, a *Avianella paragallinarum* infection model was established in a closed environment. Broilers were challenged with the virus to simulate the respiratory disease process, ensuring that the collected sound clips contained real symptoms such as coughing, making them more representative and scientifically sound, thus solving the problem of insufficient data authenticity in previous collections. Audio data was continuously collected throughout the entire period—before, during, and after infection—using a data acquisition module. Each hourly audio clip was saved as a separate file to ensure the integrity of the time series. Simultaneously, multi-area microphones were used for synchronous acquisition to cover individual sounds from different spatial distributions. The audio files were in WAV format. The entire period lasted 10 days, collecting 201 hours of valid data, which was then analyzed in chronological order. The data acquisition module consisted of six components: a microphone, a network video recorder, a switch, a computer, a power system, and an external hard drive.

[0035] Then, the acquired audio was resampled and denoised, with all audio resampled to 16kHz. Specifically, spectral gating was used to suppress background noise, retaining only the active acoustic segments of the chickens; the audio was processed into frames with a frame length of N=512 and a frame shift of 256. High-frequency response was enhanced by pre-emphasis, Hamming windows were used to smooth the edges of each frame to reduce spectral leakage, and a dual-threshold algorithm of short-time energy and short-time zero-crossing rate was used for endpoint detection to accurately extract target sound segments (sneezes, chicken calls, etc.) of the active segments. A structured raw dataset was constructed for subsequent analysis and model training.

[0036] The expression for pre-emphasis is:

[0037] ;

[0038] In the formula, This represents the sampled value of the pre-emphasized signal at time n. For the original input signal at the th The sampled value at time; It is the pre-emphasis coefficient, and its value usually ranges from 0.9 to 1.0.

[0039] The expression for calculating the window function is:

[0040] ;

[0041] In the formula, For the first Frame in Sampling points at location; This is the window function value applied to this location; assuming the window function is the same as the frame length, The value ranges from 1 to In this embodiment, since the chicken's calls have a high frequency, a Hamming window is used to effectively suppress spectral leakage caused by truncation, resulting in more accurate spectral estimation. The expression is as follows:

[0042] , ;

[0043] In the formula, For window functions, The length of the window function.

[0044] The expression for Short-Time Energy (STE) is:

[0045] ;

[0046] In the formula, It is the first The short-time energy of a frame (the sum of the squares of all amplitudes within the frame); For the first The audio amplitude of each sampling point; This represents the total number of sampling points in one frame.

[0047] The expression for the short-time average zero-crossing rate (STZCR) is:

[0048] ;

[0049] In the formula, This is a sign function that returns the sign of the input signal. The frame length.

[0050] (2) Feature parameter extraction and CNN+SE model training:

[0051] This embodiment fully considers the changes in the frequency characteristics of chickens' voices after they are infected with respiratory diseases. It selects several representative acoustic feature parameters, including: the overall envelope, reflecting the speech signal; Mel frequency cepstral coefficients (MFCCs), suitable for identifying brief, sudden sounds and representing the energy distribution of each frequency band; Mel subband energy distributions (MFENs), supplementing the detailed frequency features that MFCCs cannot fully characterize, reflecting the energy distribution of different pitches; a chroma spectrum, which helps capture changes in spectral distribution and enhances the detection of differences between local frequency bands; and spectral contrast, suitable for identifying changes in the intensity of the mid-frequency bands in coughing and evaluating the harmonic component extraction of the ratio of the fundamental frequency to higher harmonics in the chicken's cough. These five types of features are concatenated in matrix form to construct a multi-dimensional structured feature map, which is used as input to the CNN+SE model to improve the model's generalization ability.

[0052] The CNN+SE model training includes a local feature extraction network based on CNN and a channel attention module. The local feature extraction network consists of convolutional layers (Conv2D), batch normalization (BatchNorm), and an activation function (ReLU). The channel attention module enhances the perception of channel weights. During model training, after two convolutional layers and batch normalization and activation function processing, max pooling is used to compress the feature dimension. Then, two fully connected layers are used for audio segment classification, outputting the probability distribution of three types of sounds (sneezing / chicken crowing / other), achieving behavioral frequency recognition statistics for hourly audio. Figure 2 As shown.

[0053] (3) Early warning of chicken diseases:

[0054] Each hourly audio segment is divided into multiple short time intervals (e.g., one segment per second), and each segment is fed into a trained CNN-SE model for classification and recognition, outputting three-category labels: sneezing, clucking, and other. The entire hourly audio segment is then traversed and statistically analyzed. The total number of sneezes, the total number of cluckings, and other irrelevant sounds are counted according to the category labels. Combined with the actual number of chickens in each hourly segment, the unit frequency (times / chicken / hour) is calculated.

[0055] For each hourly audio segment, six acoustic feature indicators are extracted: for intensity, mean RMS energy, RMS energy standard deviation, maximum loudness, and maximum amplitude; for spectrum, spectral centroid and spectral flow. Simultaneously, frequency parameters such as rooster crowing frequency and sneezing frequency, based on the model training output, are combined, resulting in a total of eight acoustic indicators. This embodiment uses Z-score normalization to normalize these eight acoustic indicators for time-series consistency analysis, as detailed below:

[0056] ;

[0057] In the formula, Indicates the first Time of the first Original values ​​of each parameter. , These represent the mean and standard deviation of the corresponding parameters, respectively.

[0058] Based on eight normalized acoustic indicators, a weighted scoring function is constructed to calculate the comprehensive score for each hour, which facilitates subsequent warning level determination, as follows:

[0059] ;

[0060] In the formula, No. The overall score for the hour; The Z-score represents the frequency of sneezes; Represents standardized acoustic parameters; This represents the weight of the corresponding parameter, and satisfies... .

[0061] In this embodiment, the collected audio is manually labeled according to the time point of the infection. Audio before infection is labeled as "healthy" and audio after infection is labeled as "infected". The comprehensive score of the whole time series is used to analyze the optimal dividing point through the maximum value of the healthy period plus margin and ROC curve to determine the three-level warning threshold.

[0062] The method of maximum value plus standard deviation margin during the healthy period assumes that the fluctuation of the score S(t) is within the normal range during the healthy period. To avoid false alarms for occasional fluctuations, a threshold is set. The expression is as follows:

[0063] ;

[0064] In the formula, This represents the maximum overall score for all audio files marked as "healthy" per hour. This represents the standard deviation of the overall score for all audio data labeled as "healthy"; This is a margin factor (usually set to 0.5–1.0) used to adjust the fault tolerance of the threshold.

[0065] When there are clear labels (i.e., whether certain hourly data indicates an infection status), an ROC curve can be constructed, and the optimal cutoff point can be selected based on the maximum Youden index. ,as follows:

[0066] ;

[0067] In the formula, The true rate represents the proportion of infected samples that are correctly identified. This represents the false positive rate, which indicates the proportion of healthy samples that are mistakenly identified as infected.

[0068] To combine the advantages of both methods, this embodiment employs a weighted fusion strategy to obtain the final comprehensive threshold. ,as follows:

[0069] ;

[0070] In the formula, This indicates the degree of confidence in the ROC curve; a higher value indicates a stronger bias towards the model's optimal classification value. ∈[0.6,0.8].

[0071] In actual production, a preliminary model is established and initial thresholds are determined through animal experiments, which can then be deployed as the first-generation version in production. Based on the first-generation version, S(t) sequences from the past 7 days or 24 hours are collected, and after excluding red alarm data that has been identified as abnormal, the remaining data are recalculated. This is used to correct the initial threshold; if the data acquisition system provides feedback (such as the time point of manual diagnosis, red alarm data that has been identified as abnormal, etc.), then an ROC can be constructed and recalculated. Recalculate the threshold and Then, the current threshold is automatically updated using a weighted average and used for the next stage of warning logic judgment, thus repeating the cycle. This sustainable threshold update mechanism can dynamically adjust the threshold based on actual operational data, improve long-term adaptability, reduce maintenance costs, and achieve self-learning and evolvable disease monitoring capabilities.

[0072] In addition, this embodiment also sets up a three-level warning level discrimination method. Based on the calculated comprehensive threshold, the comprehensive score is divided into three colors: green, yellow, and red. Green indicates normal, yellow indicates warning, and red indicates abnormal, realizing quantitative and hierarchical management of the chicken flock's health status. The low threshold is defined as the starting point for yellow, and the comprehensive threshold is defined as the starting point for red, as follows:

[0073] The expression for the low threshold is:

[0074] ;

[0075] In the formula, This is a sensitivity adjustment item (e.g., 0.3). (This is used to prevent the system from being "too conservative" or "too sensitive").

[0076] Warning level determination rules:

[0077] when When it is green, it indicates normal; when When it is yellow, it indicates a warning; when When this occurs, it is displayed in red, indicating an abnormality. This method can not only detect respiratory abnormalities in chicken flocks in advance, but also has the advantages of strong stability, high real-time performance, and good adaptability, making it suitable for long-term deployment and online operation in actual poultry farms.

[0078] In addition, the model deployment structure of this invention is lightweight and supports embedding recognition and early warning algorithms into edge devices (such as Jetson Nano or industrial main control boards). It can independently complete sound acquisition, processing and early warning judgment in chicken farms without cloud connection, which is convenient for actual production applications.

[0079] Therefore, the broiler sneezing and respiratory disease early warning method based on CNN-SE parameter fusion adopted in this invention can effectively identify abnormal respiratory sounds, has stronger discrimination ability, and meets the actual needs of early warning of poultry respiratory diseases. At the same time, by combining the maximum value + standard deviation margin method of the healthy period and the optimal point method of ROC curve based on infection label, a three-level early warning threshold is set, which significantly enhances the sensitivity and stability of the system and avoids excessive alarms and missed alarms.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion, characterized in that, Includes the following steps: S1. An infection model was established by challenging with Avianella paragallinarum, and full-time-series audio data was collected. The collected audio data was then preprocessed. S2. Determine the acoustic feature parameters and concatenate them in matrix form to construct a multi-dimensional structured feature map as the model input. At the same time, use the CNN+SE model to construct a three-classification model to identify sneezing, crowing and other sound types in chicken sounds. S3. Based on the trained CNN+SE model, classify and identify the audio data for each hour and obtain acoustic features of intensity and spectrum. Combine this with the frequency of chicken calls and sneezing to calculate the comprehensive score for each hour. S4. By using the maximum value and standard deviation margin method during the healthy period and the optimal point method based on the ROC curve of infection label, a comprehensive threshold is determined and a three-level early warning level discrimination method is set to provide early warning for chicken diseases.

2. The method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion according to claim 1, characterized in that, In S1, the preprocessing operations include resampling and denoising. The denoising process includes suppressing background noise using spectral gating, performing frame segmentation on the audio, enhancing the high-frequency response through pre-emphasis, smoothing the edges of each frame using a Hamming window, and extracting the acoustic signal of the active segment using short-time energy and short-time zero-crossing rate.

3. The method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion according to claim 1, characterized in that, In S2, the acoustic characteristic parameters include the overall envelope, Mel frequency cepstral coefficients, Mel subband energy distribution, chromaticity spectrum, and spectral contrast.

4. The method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion according to claim 1, characterized in that, In S2, the CNN+SE model includes a local feature extraction network and a channel attention module. The local feature extraction network is based on CNN and consists of convolutional layers, batch normalization, and activation functions. The channel attention module is used to enhance the ability to perceive channel weights.

5. The method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion according to claim 1, characterized in that, In S3, intensity-related acoustic features include average root mean square energy, root mean square energy standard deviation, maximum loudness, and maximum amplitude; frequency-related acoustic features include spectral centroid and spectral flow.

6. The method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion according to claim 1, characterized in that, S3 includes normalization processing of acoustic features of intensity and frequency spectrum, as well as chicken call and sneeze frequencies.

7. The method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion according to claim 1, characterized in that, S4 includes determining the first threshold using the maximum value and standard deviation margin method during the healthy period, determining the second threshold using the optimal point method based on the ROC curve of the infection label, and determining the comprehensive threshold using a weighted average, as shown in the following expression: ； in, ； ； In the formula, As a comprehensive threshold, The first threshold, The second threshold, To determine the degree of confidence in the ROC curve, This represents the maximum overall score for each hour of audio recordings during the health period. This represents the standard deviation of the overall score for each hour of audio recordings taken during the healthy period. This is the margin coefficient. Indicates the true rate, This indicates the false positive rate.

8. The method for early warning of sneezing respiratory disease in broilers based on CNN-SE parameter fusion according to claim 1 or claim 7, characterized in that, S4 includes updating thresholds using monitoring data during the actual production process.

Images